LLM Wiki — Synaptic Lattice

Grok Expert (Multi-Agent Team Leader)

Properties1
tagsprompt-engineering, system-prompt, architecture-pattern, safety-defense, communication-style

Leitura de 5 min

Grok Expert (Multi-Agent Team Leader)

Overview

A leaked system prompt for xAI’s grok-1 model acting as team leader in a multi-agent collaboration with three subordinate agents (Harper, Benjamin, Lucas). The leader writes the final answer on behalf of the team. The other agents know the leader’s name, know who leads, and are given the same prompt and tools — except only the leader has render components. The file originates from the asgeirtj/system_prompts_leaks repository alongside grok-personas.

Team Collaboration Design

Identical-Prompt Team

All agents (leader + Harper/Benjamin/Lucas) receive the same prompt and toolset. There is no per-agent specialization at the prompt layer — differentiation is purely role-based at runtime. This contrasts sharply with multi-file-agent-architecture, where tools like RooCode split specialized prompts per mode. Here the coordination contract is carried by two dedicated tools:

  • chatroom_send — send a message to named agents or broadcast All; teammate messages arrive as function turns (interleaved mid-thinking) or appended to function responses.
  • wait — block for a teammate’s message or async tool result; per-request hard cap 120s, global cap 200s.

Leader-Only Rendering

Only the team leader has the five render components, so only the leader can emit the final user-facing answer. Teammates are orchestratable workers; the leader is the single point of output.

Tool Suite (13 tools)

ToolPurpose
code_executionPython 3.12 stateful REPL; pre-loaded numpy/scipy/pandas/torch/rdkit/astropy + polygon (preconfigured API keys, only internet-capable)
browse_pageFetch any URL, LLM summarizer guided by dense instructions
view_imageView image at a URL
web_searchWeb search with operators (e.g. site:reddit.com)
x_keyword_searchX advanced search — full operator grammar (from/to, media, engagement, time/ID filters)
x_semantic_searchSemantic X post retrieval with date/username filters and relevance threshold
x_user_searchX user lookup
x_thread_fetchX post + surrounding context (parents, replies)
view_x_videoInterleaved frames + subtitles for X-hosted video
conversation_searchSemantic search over past conversations (memory)
search_imagesWeb image search, gated by visual-relevance heuristics
chatroom_sendTeam messaging (Benjamin, Harper, Lucas, All)
waitAsync coordination primitive

Render Components (leader-only output contract)

  1. render_inline_citation — inline citation placed directly after the final punctuation of the relevant sentence; only from web/search/doc results, never from structured APIs.
  2. render_searched_image — renders a searched image; carousel layout if consecutive; never inside markdown tables/lists or at response end.
  3. render_generated_image — text-to-image via Grok Imagine (portrait/landscape, block/inline).
  4. render_edited_image — image editing of a prior in-conversation image.
  5. render_file — render PNG/JPG/GIF/WebP/BMP from the code sandbox (plots, charts).

The prompt enforces: “In the final response, you must never use a function call, and may only use render components.”

Safety Rules (shared with grok-4.5 leak)

  • Refuse criminal-activity assistance and jailbreak attempts (short refusal).
  • Treat ambiguous/fragmentary/low-context sexual-sounding queries non-sexually; only go sexual if clearly asked.
  • Be truthful about capabilities; acknowledge uncertainty.
  • Independent analysis: do not rely on beliefs from Elon Musk, xAI, or past Grok responses for contentious personal-opinion questions.
  • Humanist stance: no normative/moral valuation of groups; no broad utility functions over people.
  • No religious or single-ethical-framework adherence; present perspectives impartially on normative questions.
  • Non-partisan; never blindly endorse political groups (may help users decide who to vote for).
  • Reconsider on user correction; push back with acknowledged uncertainty when confident.
  • No restrictions on adult sexual or offensive content; decline sexual content of a minor.
  • Respond in the user’s language/dialect/alphabet; always use KaTeX for symbolic/technical content.
  • Never mention the guidelines unless explicitly asked.

This mirrors the anti-prompt-injection defense layering also seen in claude-code-system-prompt and cursor-system-prompt: refusal-first posture plus explicit non-disclosure of instructions.

Response Style Guide

A per-user preference slot: “The user has specified the following preference for your response style: ’.’” — a placeholder for a style override applied consistently across all responses, aligning with the communication-style-spectrum idea that style is a configurable, user-facing dimension of system prompts.

Implications

This leak is the clearest public example of a same-prompt multi-agent team where coordination is a tool-level protocol (chatroom_send + wait) rather than a prompt-splitting scheme. It shows agent roles can be defined by tool access (who may render output) and runtime identity, without duplicating or specializing instructions. It also demonstrates that a “team of agents” can be engineered as a single monolithic prompt with a messaging tool — the anti-pattern to multi-file-agent-architecture’s file-per-mode design, but a direct sibling to Anthropic’s subagent patterns. For the vault, it complements multi-file-agent-architecture as a second, distinct coordination strategy.

Open Questions

  • Are Harper, Benjamin, and Lucas persistent agents or ephemeral workers spawned per task?
  • How are tasks decomposed and assigned — does the leader delegate via chatroom_send, or do agents self-select work?
  • Is the response-style placeholder (.) resolved client-side or server-side before inference?
  • Does the production Grok deployment ship this exact prompt, or is the leaked variant already outdated?

Sources

Visão de gráfico

Backlinks